At a press conference yesterday, AMD proved to any doubting onlookers that they are done for. The subtext was that the company is going to make ARM64 SoCs in 2014 and put them in servers alongside x86-64 SoCs.
Lets start with a little history on the two things that make today’s debacle possible, SeaMicro and ARM. SemiAccurate was the first to write up the synergies between them starting over two years ago. As SemiAccurate said in 2010, “One interesting side note is that the ASIC talks to the CPU’s chipset via a standard PCIe interface. Seamicro says that the ASIC is CPU independent, so you could theoretically put in ARM, PPC, full power x86, or any other CPUs you wanted.” Step forward 18 months and you have the full power x86 parts in the box and ARM cores coming. Who could have foreseen that eventuality?
Then there is the whole ARM tie up, something that SemiAccurate again exclusively wrote up over a year ago. We first told you how the two mirrored each other’s views, that they were going for a common bus, and could mix and match cores in a modular fashion. In fact, we laid it out almost exactly in the summer of 2011. Other than the A15 intercept being one generation off, our explanation is dead on. “In summary, while neither side has explicitly said this directly, it is painfully clear that AMD and ARM are going to make a common on-chip interface and interconnection. You can take a Bobcat chip, yank the GPU, and slap a Mali or Imagination block in. Want a quad-core A15 that runs ATI demos? A shoelace tip controller that bootstraps a 6970? How about a Bulldozer or Trinity that uses that custom accelerator block that Google or Facebook is probably working on? Crypto blocks that are common to ARM and AMD? No problem. Get the picture now?”
Back to the present, in today’s press conference AMD CEO Rory Reed showed the depths of how poorly he understands what AMD is doing. A simple phrase that will probably be spun as marketing hype was that Reed called the SeaMicro Freedom Fabric (FF) a “supercompute fabric”. This is repeated in the press release, “The first ARM technology-based AMD Opteron™ processor is targeted for production in 2014 and will integrate the AMD SeaMicro Freedom™ supercompute fabric, the industry’s premier high-performance fabric.” The problem with this is that it is technically wrong.
FF is many things, but supercomputing is not one of them. It is an amazing backbone tech for a big box of shared nothing (BBOSN) machines, but it is fundamentally unsuited for supercomputing tasks. Anyone who had the vaguest clue about the product would understand this and not make that basic mistake. It is like calling a bicycle a car because both have similar privileges on some roads, analysts might be mislead, but that does not make it true. It is extremely worrisome that Mr Reed does not understand the tech on this level, it isn’t that complex and it fundamentally affects his company.
AMD went on in their traditional fashion, lots and lots of superlatives around two bullet points, then avoided any useful information or details. They are making an ARM64 ISA SoC that has the FF in it, and that there will be SeaMicro boxes with both x86 and ARM CPU cores in them. AMD danced around the issue of cores, and would not answer the questions about the core itself being an ARM design or an AMD custom core. Luckily Anand was able to get that info, the core is not custom.
Putting the FF on die means that power per socket, or power per node in this case, drops substantially. This is a good thing. As we said years ago, the FSA/HSA common bus will allow AMD x86 and ARM cores to be mix and matched fairly easily and quickly. Better yet, you can add 3rd party IP too, and there will be a plethora of that from the ARM side too. So far, it sounds great doesn’t it?
One big problem is that AMD is late, very late to a rather crowded game. This whole strategy seems to be one of closing of the barn doors after the horse has already left. There are several other companies that are not just doing this, but have products on the market now with ARM cores in them. Calxeda and Marvell among others have CPUs out now, and you can buy the servers from Dell, HP, Boston, and Penguin among others. This is released product, AMD has a vague timetable.
Worse yet, there is the bigger problem of the fabric itself relative to the competition. Calxeda has a fabric that is far more configurable than what AMD has in their FF. This is not to say that FF is bad, simple, or anything of the sort, quite the opposite, it is brilliant at what it does. Calxeda has different design goals from SeaMicro, and the end result is a very different product. To make matters worse, Rory Reed’s comments about the supercompute fabric could be done by Calxeda but it can’t be done by AMD/SeaMicro. Once again, this isn’t to knock SeaMicro, it just was designed to address a different market.
So AMD is facing competition on the market now with more flexible designs, IP that AMD doesn’t have, and designs that AMD can’t match. The cores AMD is fielding are vanilla ARM64/V8 cores that anyone with a checkbook can buy as well, there is no magic there. Worse yet, there are others stepping up to the plate with full custom ARM64/V8 cores that should savage the ARM reference core on energy use, peak performance, and sometimes both.
That brings us to the most damning point, cores. As one of the guests on the panel at the AMD press conference today said, the cloud is ISA agnostic. It might have been the Facebook representative but we didn’t catch exactly who he was because the webcast had technical problems. Paraphrased he said that the cloud was great because everything is open sourced, the tools are there, and the end user doesn’t care what their code runs on. We totally agree with this outlook, and have been pointing it out for years in our consulting work.
For AMD, this is a big problem. Why? The space they are looking at, dense servers and microservers, live and die on two basic metrics. Those are performance per watt and absolute single threaded performance. If AMD does their ‘ambidextrous’ strategy, they will have the best of neither world. In fact, they won’t even be close on either metric.
The Kabini core may be great but it isn’t even in the same ballpark as an Intel iSomethingmeaningless core for single threaded performance or performance per watt. It may fill some really lucrative device niches well and do exactly what it was meant to do, but dense servers are not what it was designed for. The Bulldozer/Piledriver core is closer to Intel on single threaded benchmarks, but woefully behind even Kabini/Jaguar on performance per watt.
Moving on to the ARM cores, AMD has a customer core that is, as we mentioned earlier, off the shelf purchasable. The rest of the uncore may be good, but AMD has absolutely zero experience in making low power ARM uncores. By the time they have a product sampling, the competition will be on generation 3 or more, plus tons of field data from customers. Other competitors will have custom cores as well, each tailored toward the market better than the generic one AMD is buying. AMD is starting out late and behind.
The Freedom Fabric is of course dead on tailored for some of the microserver and dense server markets AMD is aiming at, and it is among the best in class. Calxeda has one that is likely just as energy efficient and its flexibility allows it to address many more markets. On top of Calxeda, there are unquestionably others that have not broken cover yet lurking in the wings, and who knows what they have in store?
Then there is Intel, they have iSomethingmeaningless cores for single threaded work, Centerton, shipping now actually, for microservers, and can license the same ARM64/V8 core as AMD if it wants to. All of this is built on a process at least two full nodes ahead of what AMD has access to on the best of days. If Intel wants to, they can crush AMD on all fronts with relative ease.
In short, AMD has a first class fabric, then, well, hey do you hear crickets too? Their cores are good, but not for the job at hand, and their design expertise is in areas not directly applicable to the task being undertaken. They are good on one front, average at best on the rest. Some of the competition is best in class on all fronts, has generations of expertise doing what AMD wants to do, and in some cases has technology that AMD simply never will.
All in all, this is why we were hoping AMD had something amazing up their sleeve. Instead, the company announced what is essentially a purchased off the shelf IP block that is not close to the best of their competition. AMD’s senior management was put in place to chase an untenable dream over the protests of someone who understood why it would fail. He is gone. The new order doesn’t understand the fabric.S|A
Latest posts by Charlie Demerjian (see all)
- SemiAccurate has Skylake-SP die shots, sizes, and more - Jul 21, 2017
- Skylake-SP has a diverged core - Jul 19, 2017
- Intel’s Purley platform architecture is a step forward - Jul 13, 2017
- Intel launches Purley aka metal Xeons - Jul 11, 2017
- AMD’s Epyc has lots of connections - Jun 26, 2017